LoginSign Up
← Back to all sparks
O

Ollama

AI-ASSISTANTS
Velocity6.3

Run and manage large language models locally on your own machine.

ollama.com ↗

Ollama integrates with ChatGPT Desktop as a local backend while the v0.34.x RC cycle hardens OpenAI API compatibility.

local-llmopenai-compatchatgpt-desktopmlxapple-siliconmodel-serving
◆Current state
Ollama is in an active RC cycle for v0.34.0/0.34.1, with the defining move being v0.34.0-rc0's integration with ChatGPT Desktop — local Ollama instances can now serve as a backend for OpenAI's own desktop app. The RC builds since have focused on hardening the OpenAI API compatibility layer: named function outputs, Codex agent message handling, web search response finalization, and proxy fixes for the ChatGPT integration. Separately, the MLX engine gained MoE global scaling support, broadening the range of large open-weight models that run well on Apple Silicon.
◆Where it's heading
Ollama is evolving from a standalone local model server into the preferred local runtime behind OpenAI-native tooling. The ChatGPT Desktop integration is the clearest signal: rather than competing for users with a distinct UX, Ollama is becoming infrastructure that feeds existing interfaces. Continued OpenAI API compatibility work and MLX engine investment point to deepening the Apple Silicon story and expanding tool-call and agent protocol coverage.
◆Prediction
The stable v0.34.0 will formalize ChatGPT Desktop as a documented integration target. v0.35 will likely close remaining OpenAI API gaps — streaming tool calls, the Responses API surface — and potentially add Windows-native ChatGPT Desktop support if the integration pattern holds.

◆Recent moves

  1. 1d ago

    v0.34.1-rc2: API: Deprecate typical_p (#18448)

    View source ↗
  2. 1d ago

    v0.34.1 RC1: Docker MLX Build Context Fix

    A build context fix adding the MLX patch to Docker builds — internal infrastructure change with no user-visible effect outside Docker-based Ollama deployments.

    View source ↗
  3. 1d ago

    v0.34.1 RC0: MLX Engine Adds MoE Global Scale Support

    MLX version bump brings ModelOpt global scaling for MoE (Mixture-of-Experts) models, expanding which large open-weight models run efficiently on Apple Silicon. MoE architectures (used by models like DeepSeek and Mixtral variants) previously had no MLX-side global scale support, limiting quantization accuracy.

    View source ↗
  4. 6d ago

    v0.34.0 RC5: Named Function Outputs in OpenAI Compatibility

    Standalone named function outputs now work in the OpenAI compatibility layer, filling a gap for callers that use structured tool responses with explicit output names — a pattern common in agent frameworks that map tool call results to named slots.

    View source ↗
  5. 6d ago

    v0.34.0 RC4: Proxy Namespace Command Fix

    A proxy-layer fix normalizing namespaced commands in Full Access mode — an internal routing correction with no visible effect for standard API consumers.

    View source ↗
  6. 7d ago

    v0.34.0 RC3: Codex Agent Message Compatibility

    The OpenAI compatibility layer now accepts plaintext-labeled Codex agent messages, widening protocol support for callers using OpenAI's Codex agent format. This is part of the broader hardening of the ChatGPT Desktop integration introduced in v0.34.0-rc0.

    View source ↗